Improved Closed-Form Prediction Intervals for

Binomial and Poisson Distributions

K. Krishnamoorthy
a
and Jie Peng
b
a
Department of Mathematics, University of Louisiana at Lafayette, Lafayette, LA 70504
b
Department of Finance, Economics and Decision Science, St. Ambrose University, Davenport, Iowa 52803
The problems of constructing prediction intervals for the binomial and Poisson distributions are con-
sidered. Available approximate, exact and conditional methods for both distributions are reviewed and
compared. Simple approximate prediction intervals based on the joint distribution of the past samples and
the future sample are proposed. Exact coverage studies and expected widths of prediction intervals show
that the new prediction intervals are comparable to or better than the available ones in most cases. The
methods are illustrated using two practical examples.
Key Words: Clopper-Pearson approach; Conditional approach; Coverage probability; Hypergeomet-
ric distribution.
-
a
Corresponding author. Tel.: +1 337 482 5283; fax: +1 337 482 5346.
E-mail address: krishna@louisiana.edu (K. Krishnamoorthy).
2
1. Introduction
In many practical situations, one needs to predict the values of a future random variable based
on the past and currently available samples. Extensive literature is available for constructing predic-
tion intervals (PIs) for various continuous probability distributions and other continuous models such
as linear regression and one-way random models. Applications of prediction intervals (PIs) based on
continuous distributions are well-known; for examples, the prediction intervals based on gamma dis-
tributions are often used in environment monitoring (Gibbons, 1987), and normal based PIs are used
in monitoring and control problems (Davis and McNichols, 1987). Other applications and examples
can be found in the book by Aitchison and Dunsmore (1980). Compared to continuous distributions,
results on constructing PIs for discrete distributions are very limited. Prediction intervals for a discrete
distribution are used to predict the number of events that may occur in the future. For example, a
manufacturer maybe interested in assessing the number of defective items in the future production
process based on available samples. Faulkenberry (1973) provided an illustrative example in which the
number of breakdowns of a system in a year follows a Poisson distribution, and the objective is to
predict the number of breakdowns in a future year based on available samples. Bain and Patel (1993)
have noted a situation where it is desired to construct binomial prediction intervals (see Section 4).
The prediction problem that we will address concerns two independent binomial samples with the
same success probability p. Given that X successes are observed in n independent Bernoulli trials,
we like to predict the number of successes Y in another m independent Bernoulli trials. In particular,
we like to nd a prediction interval [L(X; n, m, ), U(X; m, n, )] so that
P
X,Y
(L(X; n, m, ) Y U(X; m, n, )) 1 2.
As the conditional distribution of one of the variables given the sum X + Y does not depend on the
parameter p, prediction intervals can be constructed based on the conditional distribution. Thatcher
(1964) noted that a PI for Y can be obtained from the conditional distribution of X given X +Y = s,
which is hypergeometric (with sample size s, number of defects n, and the lot size n + m) and does
not depend on p. Thatchers method is similar in nature to that of Clopper-Pearsons (1934) ducial
approach for constructing exact condence intervals for a binomial success probability. Kn usel (1994)
has pointed out that the acceptance regions of two-sample tests (one-sided tests for comparing success
probabilities of X and Y ) are exact PIs for Y . We found that the PIs for the binomial case given in
Kn usel (1994), and the one described in Aitchison and Dunsmore (1980, Section 5.5) are indeed the
same as the exact PIs that can be obtained by Thatchers (1964) method. Faulkenberry (1973) provided
a general method of constructing a PI on the basis of the conditional distribution given a complete
sucient statistic. He has illustrated his general approach for constructing PIs for the exponential and
Poisson distributions based on the conditional distribution of Y (the variable to be predicted) given
the sum X + Y . Bain and Patel (1993) have applied Faulkenberrys approach to construct PIs for
the binomial, negative binomial and hypergeometric distributions. Dunsmore (1976) has shown that
3
Faulkenberrys approach is not so quite general as claimed, and provided a counter example. Kn usel
(1994, Section 3) noted that Faulkenberrys approach is correct for the continuous case, but it is not
exact for the Poisson case.
Apart from the conditional methods given in the preceding paragraph, there is an asymptotic
approach proposed in Nelson (1982), which is also reviewed in Hahn and Meeker (1991). Nelsons
PI is based on the asymptotic normality of a standard pivotal statistic, and it is easier to compute
than the ones based on the conditional approaches. However, Wangs (2008) coverage studies and our
studies (see Figures 1b and 3) indicate that Nelsons PIs have poor coverage probabilities even for large
samples. Wang considered a modied Nelsons PI in which a factor has to be determined so that the
minimum coverage probability is close to the nominal condence level. Wang has provided a numerical
approach to nd the factor. A reviewer has brought to our attention that Wang (2010) has proposed
another closed-form PI for a binomial distribution. It should be noted that the exact and conditional
PIs were not included in Wangs (2008, 2010) coverage studies.
All the methods for the binomial case are also applicable for the Poisson case as shown in the sequel.
In this article, we propose a closed form approximate PIs based on the joint sampling approach
which is similar to the one used to nd condence interval in a calibration problem (e.g., see Brown,
1982, Section 1.2). The proposed PIs are simple to compute, and they are comparable to or better
than the available PIs in most cases. Furthermore, we show that the conditional PIs are narrower than
the corresponding exact PIs except for the extreme cases where the conditional PIs are not dened.
We also show that the new PI is included in or identical to Wangs (2010) PI.
The rest of the article is organized as follows. In the following section, we review the conditional
methods, Nelsons PIs, Wangs (2010) PIs, and present the new PIs based on the joint sampling
approach for the binomial case. The exact coverage probabilities and expected widths of the PIs are
evaluated. The results for the binomial case are extended to the Poisson distribution in Section 3.
Construction of the binomial and Poisson PIs are illustrated using two examples in Section 4. Some
concluding remarks are given in Section 5.
2. Binomial Distribution
Let X binomial(n, p) independently of Y binomial(m, p). The problem is to nd a 1 2
prediction interval for Y based on X. The conditional distribution of X given the sum X + Y = s
is hypergeometric with the sample size s, number of non-defects n, and the lot size n + m. The
conditional probability mass function is given by
P(X = x[X +Y = s, n, n +m) =
_
n
x
__
m
y
_
_
m+n
s
_ , max0, s m x minn, s.
4
Let us denote the cumulative distribution function (cdf) of X given X + Y = s by H(t; s, n, n + m).
That is,
H(t; s, n, n +m) = P(X t[s, n, n +m) =
t

i=0
_
n
i
__
m
si
_
_
m+n
s
_ . (1)
Note that the conditional cdf of Y given X +Y = s is given by H(t; s, m, n +m).
2.1 Binomial Prediction Intervals
2.1.1 The Exact Prediction Interval
Thatcher (1964) developed the following exact PI on the basis of the conditional distribution of X
given X + Y . Let x be an observed value of X. The 1 lower prediction limit L is the smallest
integer for which
P(X x[x +L, n, n +m) = 1 H(x 1; x +L, n, n +m) > . (2)
The 1 upper prediction limit U is the largest integer for which
H(x; x +U, n, n +m) > . (3)
Furthermore, [L, U] is a 12 two-sided PI for Y . Thatcher (1964) has noted that, for a xed (x, n, m),
the probability (3) is a decreasing function of U, and so a backward search, starting from m, can be
used to nd the largest integer U for which the probability in (3) is just greater . Similarly, we see
that the probability in (2) is an increasing function of L, and so a forward search method, starting
from a small value, can be used to nd the smallest integer L for which this probability is just greater
than .
The exact PIs for extreme values of X are dened as follows. When X = 0, the lower prediction
limit for Y is 0, and the upper one is determined by (3); when x = n, the upper prediction limit is m,
and the lower prediction limit is determined by (2).
2.1.2 The Conditional Predication Interval
Bain and Patel (1993) used Faulkenberrys (1973) conditional approach to develop a PI for Y . As
noted earlier, this PI is based on the conditional distribution of Y , conditionally given X + Y , and is
described as follows. Recall that the conditional cdf of Y given X+Y = s is given by P(Y t[X+Y =
s) = H(t; s, m, n + m). Let x be an observed value of X. The 1 lower prediction limit L for Y is
the largest integer so that
H(L 1; x +L, m, n +m) , (4)
and the 1 upper prediction limit U is the smallest integer for which
H(U; x +U, m, n +m) 1 . (5)
5
Furthermore, [L, U] is the 1 2 two-sided PI for Y . As noted by Dunsmore (1976), Faulkenberrys
method, in general, is not exact. We also note that the above prediction interval is not dened at the
extreme cases X = 0 and X = n.
2.1.3 The Nelson Prediction Interval
Let p =
X
n
, q = 1 p and

Y = m p. Using the variance estimate var(

Y Y ) = m p q(1 + m/n),
Nelson (1982) proposed an approximate PI based on the asymptotic result that

Y Y
_
var(

Y Y )
=

Y Y
_
m p q(1 +m/n)
N(0, 1). (6)
The resulting 1 2 PI is given by

Y z
1
_
m p q
_
1 +
m
n
_
, where z

is the 100 percentile of the

standard normal distribution. Notice that the above PI is not dened when X = 0 or n. As X assumes
these values with positive probabilities, the coverage probabilities of the above PI are expected to be
much smaller than the nominal level when p is at the boundary (Wang, 2008). To overcome the poor
coverage probabilities at the boundary, we can dene the sample proportion as p =
0.5
n
if X = 0 and
n0.5
n
if X = n. This type of adjustments is commonly used to handle extreme cases while estimating
the odds ratio or the relative risk involving two binomial distributions (e.g., see Agresti, 1999). Since
we are predicting a discrete random variable, the PI is formed by the set of integer values of Y satisfying

Y Y

m p q(1+m/n)

< z
1
, and is given by
__

Y z
1

Y (m

Y )
_
1
m
+
1
n
_
_
,
_

Y +z
1

Y (m

Y )
_
1
m
+
1
n
_
__
, (7)
where x| is the smallest integer greater than or equal to x, and x| is the largest integer less than or
equal to x.
2.1.4 Wangs Prediction Interval
Wang (2010) proposed a PI by employing an approach similar to the construction of the Wilsons
score CI for the binomial proportion p. Wangs PI is based on the result that
W(m, n, X, Y ) =
Y m p
__
X+Y +c/2
m+n+c
__
1
X+Y +c/2
m+n+c
_
m(m+n)
n
_1
2
N(0, 1), approximately, (8)
where c = z
2
1
. The 1 2 PI is the set of integer values of Y that satisfy [W(m, n, X, Y )[ < z
2
1
.
6
To write the PI explicitly, let
A = mn
_
2Xz
2
1
(n +m+z
2
1
) + (2X +z
2
1
)(m+n)
2

,
B = mn(m+n)z
2
1
(m+n +z
2
1
)
2
2(n X)[n
2
(2X +z
2
1
) + 4mnX + 2m
2
X]
+ nz
2
1
[n(2X +z
2
1
) + 3mn +m
2
],
C = 2n
_
(n +z
2
1
)(m
2
+n(n +z
2
1
)) +mn(2n + 3z
2
1
)

.
In terms of above notations, Wangs PI is given by [L|, U|], where
(L, U) =
A
C

B
C
. (9)
2.1.5 The Prediction Interval based on the Joint Sampling Approach
The PI that we propose is based on the complete sucient statistic X+Y for the binomial(n+m, p)
distribution. Dene p
xy
=
X+Y
n+m
, and var(m p
xy
Y ) = mn p
xy
q
xy
/(n + m), where q
xy
= 1 p
xy
. We
can nd a PI for Y based on the asymptotic joint sampling distribution that
m p
xy
Y
_
var(m p
xy
Y )
=
(mX nY )
_
mn p
xy
q
xy
(n +m)
N(0, 1). (10)
To handle the extreme cases of X, let us dene X to be 0.5 when X = 0, and to be n.5 when X = n.
The 1 2 prediction interval is the set of integer values of Y satisfying
(mX nY )
2
mn p
xy
q
xy
(n +m)
< z
2
1
. (11)
Treating as if Y is continuous, we can nd the roots of the quadratic equation
(mXnY )
2
mn p
xy
q
xy
(n+m)
= z
2
1
as
_

Y
_
1
z
2
1
m+n
_
+
mz
2
1
2n
_
z
1
_

Y (m

Y )
_
1
m
+
1
n
_
+
m
2
z
2
1
4n
2
_
1 +
mz
2
1
n(m+n)
_ = (L, U), say, (12)
where

Y = mX/n for X = 1, ..., n 1; it is .5m/n when X = 0 and (n .5)m/n when X = n. The
1 2 PI, based on the roots in (12), is given by [L| , U|] . We refer to this PI as the joint sampling
- prediction interval (JS-PI).
2.2 Comparison Studies and Expected Widths of Binomial Prediction Intervals
The conditional PI is either included in or identical to the corresponding exact PI except for the
extreme cases where the former PI is not dened (see Appendix A for a proof). In order to study the
coverage probabilities and expected widths of the conditional PI, we take it to be the exact PI when
7
X = 0 or n. Specically, the conditional PI is determined by (4) and (5) when X = 1, ..., n 1, and
by (2) or (3) when X = n or 0. Between our JS-PI and Wangs (2010) PI, the former is included in or
identical to Wangs (2010) PI (see Appendix B for a proof). There are several cases where our JS-PI
is strictly shorter than the Wang PI as shown in the following table.
95% PIs
(n, m) X Wangs PI JS-PI
(40, 20) 16 [3, 13] [4, 13]
33 [12, 20] [12, 19]
(100,50) 10 [1, 11] [1, 10]
(300,100) 18 [1, 12] [2, 11]
(1000,300) 278 [67,101] [67, 100]
(3000,100) 212 [2, 12] [3, 12]
Note that for all the cases considered in the above table, our JS-PI is shorter than Wangs PI by only
one unit.
To compare PIs on the basis of coverage probabilities, we note that, for a given (n, m, p, ), the
exact coverage probability of a PI [L(x, n, m, ), U(x, n, m, )] can be evaluated using the expression
n

x=0
m

y=0
_
n
x
__
m
y
_
p
x+y
(1 p)
m+n(x+y)
I
[L(x,n,m,),U(x,n,m,)]
(y), (13)
where I
A
(.) is the indicator function. For a good PI, its coverage probabilities should be close to the
nominal level.
We evaluated the coverage probabilities of all ve PIs using (13) for some sample sizes and condence
coecient 0.95. In order to display the plots clearly, we plotted the coverage probabilities of Wangs
(2010) PIs in (9) and those of the JS-PIs (12) in Figure 1a. Even though we knew that the JS-PI
is contained in Wangs PI, the coverage probabilities of these two PIs should be evaluated because
both are approximate. It is clear from these plots that both PIs could be liberal, but their coverage
probabilities are barely below the nominal level 0.95. As expected, for all sample size congurations
considered, the coverage probabilities of Wangs PIs are greater than or equal to those of the JS-PIs.
Thus, Wangs PIs are more conservative than the JS-PIs for some parameter space, especially when p
is away from 0.5. As a result, Wangs PIs are expected to be wider than the JS-PIs. In view of these
comparison results, we will not include Wangs PIs for further comparison study in the sequel.
The coverage probabilities of the PIs based on the exact, conditional, Nelsons PIs and joint sampling
methods are plotted in Figure 1b. We rst observe from these plots that all the PIs, except the Nelson
PI, are conservative when p is close to zero or one. The exact PIs are too conservative even for large
samples; see the coverage plots for n = m = 50, n = 50, m = 20 and n = 100, m = 20. As expected, the
8
0.9
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
0 0.2 0.4 0.6 0.8 1
n = m = 15
Wang
JS
0.9
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
0 0.2 0.4 0.6 0.8 1
n = 20, m = 10
Wang
JS
0.9
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
0 0.2 0.4 0.6 0.8 1
n = 50, m = 50
Wang
JS
0.9
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
0 0.2 0.4 0.6 0.8 1
n = 50, m = 20
Wang
JS
0.9
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
0 0.2 0.4 0.6 0.8 1
n = 100, m = 50
Wang
JS
0.9
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
0 0.2 0.4 0.6 0.8 1
n = 100, m = 20
Wang
JS
Figure 1a. Coverage probabilities of 95% binomial PIs as a function of p
9
conditional PIs are less conservative than the exact PIs for all the cases considered. Furthermore, the
coverage probabilities of the conditional PIs are seldom fall below the nominal level 0.95; see the plots
for n = m = 15 and n = m = 50. Nelsons PIs are in general liberal, especially when the values of the
parameter are at the boundary. Overall, the exact PIs are too conservative, and Nelsons PIs are liberal
even for large samples. The conditional PI and the new one based on the joint sampling approach (JS-
PI) are comparable in some situations (see the plots for n = m = 15 and n = m = 50), and the latter
exhibits better performance than the former in other cases. In particular, in the most common case
where the past and current samples are larger than the future sample, the new PIs are better than the
corresponding conditional PIs (see the plots for the cases n = 50, m = 20 and n = 100, m = 20).
As the conditional PI and the new PI are the ones that control the coverage probabilities satisfac-
torily, we compare only these PIs with respect to expected widths. We evaluate the expected widths
of these PIs for the sample size congurations for which coverage probabilities were computed earlier.
The plots of the expected widths are given in Figure 2. These plots indicate that the expected widths of
the new PIs are comparable to or smaller than those of the conditional PIs for all the cases considered.
In general, the new prediction interval is preferable to the conditional PI with respect to coverage
properties and expected widths.
3. Poisson Distribution
Let X be the total counts in a sample of size n from a Poisson distribution with mean . Note that
X Poisson(n). Let Y denote the future total counts that can be observed in a sample of size m
from the same Poisson distribution so that Y Poisson(m). Note that the conditional distribution
of X, conditionally given X + Y = s, is binomial with number of trials s and the success probability
n/(n + m), binomial(s, n/(m + n)), and the conditional distribution of Y given the sum X + Y is
binomial(s, m/(m + n)). Let us denote the cumulative distribution function of a binomial random
variable with the number of trials N and success probability by B(x; N, ).
3.1 The Exact Prediction Interval
As noted earlier, the exact PI is based on the conditional distribution of X given X +Y . Let x be
an observed value of X. The smallest integer L that satises
1 B(x 1; x +L, n/(n +m)) > (14)
is the 1 lower prediction limit for Y . The 1 upper prediction limit U is the largest integer for
which
B(x; x +U, n/(n +m)) > . (15)
For X = 0, the lower prediction limit is dened to be zero, and the upper prediction limit is determined
by (15).
10
0.9
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
0 0.2 0.4 0.6 0.8 1
n = m = 15
exact
cond.
Nelson
..
..
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..
.
.
.........
..
..
...
.
...
.
...
..
.
...
..
....
..
.....
...
....
.............
.....
.............
....
...
.....
..
....
..
...
.
..
...
.
...
.
...
..
..
.........
.
.
..
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..
..
.
JS

0.9
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
0 0.2 0.4 0.6 0.8 1
n = 20, m = 10
exact
cond.
Nelson
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..
...
.......
.....
..........................
.....
.................................................................
.....
..........................
.....
.......
...
..
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
JS

0.9
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
0 0.2 0.4 0.6 0.8 1
n = m = 50
exact
cond.
Nelson
.
.
.
.
.
.
.
.
.
.
....
.
.
.
.
.
.
.
.
.
.
..
.
..
.
....
.
..
....
..
....
...
..................
...........
.........................................
...........
..................
...
....
..
....
..
.
....
.
..
.
..
.
.
.
.
.
.
.
.
.
.
....
.
.
.
.
.
.
.
.
.
.
.
JS

0.9
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
0 0.2 0.4 0.6 0.8 1
n = 50, m = 20
exact
cond.
Nelson
.
.
.
.
.
.
.
.
.
.
..
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..
.
..
..
...
..
...
....................
......
.................................................................
......
....................
...
..
...
..
..
.
..
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..
.
.
.
.
.
.
.
.
.
.
.
JS

0.9
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
0 0.2 0.4 0.6 0.8 1
n = 100, m = 50
exact
cond.
Nelson
.
.
.
.
.
..
.
.
.
.
.
..
...........
.............
......
.....
..................
......
.........
................
..
.................
.........
......
..................
.....
......
.............
...........
..
.
.
.
.
.
..
.
.
.
.
.
.
JS

0.9
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
0 0.2 0.4 0.6 0.8 1
n = 100, m = 20
exact
cond.
Nelson
.
.
..
.
.
.
.
.
.
.
.
.
.
..
..
...
............
...
............
.....
...............................
.....
.....................
.....
...............................
.....
............
...
............
...
..
..
.
.
.
.
.
.
.
.
.
.
..
.
.
.
JS

Figure 1b. Coverage probabilities of 95% binomial PIs as a function of p

11
4
5
6
7
8
9
10
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
n = m = 15
cond.
JS
2
2.5
3
3.5
4
4.5
5
5.5
6
6.5
7
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
n = 20, m = 10
cond.
JS
4
6
8
10
12
14
16
18
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
n = m = 50
cond.
JS
2
3
4
5
6
7
8
9
10
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
n = 50, m = 20
cond.
JS
2
4
6
8
10
12
14
16
18
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
n = 100, m = 50
cond.
JS
1
2
3
4
5
6
7
8
9
10
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
n = 100, m = 20
cond.
JS
Figure 2. Expected widths of 95% binomial prediction intervals as a function of p
12
3.2 The Conditional Prediction Interval
The conditional approach is based on the conditional distribution of Y given the sum X +Y . The
1 lower prediction limit is the largest integer L for which
B(L 1; x +L, m/(m+n)) , (16)
where x is an observed value of X. The 1 upper prediction limit U is the smallest integer for which
B(U; x +U, m/(m+n)) 1 . (17)
3.3 The Nelson Prediction Interval
Let

= X/n, and var(m

Y ) = m
2

(1/n + 1/m). Nelsons approximate PI is based on the

asymptotic result that (m

Y )/
_
var(m

Y ) N(0, 1), and is given by

[L|, U|] , with [L, U] =

Y z
1

Y
_
1
m
+
1
n
_
, (18)
where

Y = m
X
n
, for X = 1, 2, ..., and is
.5m
n
when X = 0.
3.4 The Prediction Interval Based on the Joint Sampling Approach
Let

xy
=
X+Y
m+n
. As in the binomial case, using the variance estimate var(m

xy
Y ) =
mn

xy
m+n
, we
consider the quantity
m

xy
Y
_
var(m

xy
Y )
=
(mX nY )
_
mn(X +Y )
,
whose asymptotic joint distribution is the standard normal. In order to handle the zero count, we take
X to be 0.5 when it is zero. The 1 2 PI is determined by the roots (with respect to Y ) of the
quadratic equation
(m

Y )
2

xy
m(1+m/n)
= z
2
1
. Based on these roots, the 1 2 PI is given by
[L|, U|], with [L, U] =

Y +
mz
2
1
2n
z
1

Y
_
1
m
+
1
n
_
+
m
2
z
2
1
4n
2
, (19)
where

Y = m
X
n
, for X = 1, 2, ..., and is
.5m
n
for X = 0.
3.5 Coverage Probabilities and Expected Widths of Poisson Prediction Intervals
As in the binomial case, it can be shown that the conditional PIs are either included in or identical
to the corresponding exact PIs except for the case of X = 0 where the former PIs are not dened. This
can be proved along the lines for the binomial case given in Appendix A, and so the proof is omitted.
13
In order to study the unconditional coverage probabilities and expected widths of the conditional PIs,
we take them to be exact PIs when X = 0.
For a given (n, m, , ), the exact coverage probability of a Poisson PI [L(x, n, m, ), U(x, n, m, )]
can be evaluated using the expression

x=0

y=0
e
(n+m)
(n)
x
(m)
y
x!y!
I
[L(x,n,m,),U(x,n,m,)]
(y).
The coverage probabilities of all PIs, except the exact one, are plotted in Figure 3 for some values of
(n, m). The exact PI is not included in the plots because it exhibited similar performance (excessively
conservative) as in the binomial case, and it is inferior to the conditional PI in terms of coverage
probabilities and expected widths. Regarding other PIs, we see from these six plots in Figure 3 that
Nelsons PIs are often liberal. The conditional PI and the new JS-PI behave similarly when n = m,
and the former is more conservative than the latter when n > m (see the plots for (n = 5, m = 2) and
(n = 10, m = 1)). Both PIs could be occasionally liberal, especially when is very small.
As the conditional PI and the new one based on the joint sampling approach are the ones satisfacto-
rily control the coverage probabilities, we further compare them with respect to expected widths. The
plots of expected widths are given in Figure 4. These plots clearly indicate that the expected widths of
these two PIs are not appreciably dierent for the situations they have similar coverage probabilities.
On the other hand, for the cases where the conditional PIs are conservative, the new PIs have shorter
expected widths; see the plots of (n = 5, m = 2) and (n = 10, m = 1).
We also evaluated coverage probabilities and expected widths of PIs with condence coecients
0.90 and 0.99. The results of comparison are similar to those of 95% PIs, and so they are not reported
here. On an overall basis, the new PI is preferable to other PIs in terms of coverage probabilities and
precisions.
4. Examples
4.1 An Example for Binomial Prediction Intervals
This example is adapted from Bain and Patel (1993). Suppose for a random sample of n =
400 devices tested, x = 20 devices are unacceptable. A 90% prediction interval is desired for the
number of unacceptable devices in a future sample of m = 100 such devices. The sample proportion
of unacceptable devices is p = 20/400 = .05 and

Y = m p = 5. The normal critical value required
to compute Nelsons PI and the new PI is z
.95
= 1.645. Nelsons PI given in (7) is [1, 9]. The new
JS-PI in (12) and Wangs PI in (9) are [2, 9]. To compute the conditional PI, we can use the new
prediction limits just obtained as initial guess values. In particular, using L = 2 as the initial guess
14
0.86
0.88
0.9
0.92
0.94
0.96
0.98
1
0 1 2 3 4 5 6 7 8 9
n = m = 1
Nelson-PI
cond. PI
JS-PI

0.86
0.88
0.9
0.92
0.94
0.96
0.98
1
0 1 2 3 4 5 6 7 8 9
n = m = 4
Nelson-PI
cond. PI
JS-PI

0.86
0.88
0.9
0.92
0.94
0.96
0.98
1
0 1 2 3 4 5 6 7 8 9
n = 5, m = 2
Nelson-PI
cond. PI
JS-PI

0.86
0.88
0.9
0.92
0.94
0.96
0.98
1
0 1 2 3 4 5 6 7 8 9
n = 2, m = 5
Nelson-PI
cond. PI
JS-PI

0.86
0.88
0.9
0.92
0.94
0.96
0.98
1
0 1 2 3 4 5 6 7 8 9
n = 10, m = 1
Nelson-PI
cond. PI
JS-PI

0.86
0.88
0.9
0.92
0.94
0.96
0.98
1
0 1 2 3 4 5 6 7 8 9
n = 1, m = 10
Nelson-PI
cond. PI
JS-PI

Figure 3. Coverage probabilities of 95% prediction intervals for Poisson distributions as a function of
15
4
6
8
10
12
14
16
0 1 2 3 4 5 6 7 8 9
n = m = 1
cond. PI
JS-PI
5
10
15
20
25
30
35
0 1 2 3 4 5 6 7 8 9
n = m = 4
cond. PI
JS-PI
2
4
6
8
10
12
14
16
18
20
0 1 2 3 4 5 6 7 8 9
n = 5, m = 2
cond. PI
JS-PI
0
2
4
6
8
10
12
14
0 1 2 3 4 5 6 7 8 9
n = 10, m = 1
cond. PI
JS-PI
Figure 4. Expected widths of 95% Poisson prediction intervals as a function of
16
value, we compute the probability in (4) as H(L 1 = 1, x + L = 22, 100, 500) = 0.0446 at L = 2;
H(L 1 = 2, x + L = 23, 100, 500) = 0.1272 > .05 at L = 3. So the lower endpoint is 2. Similarly,
we can nd the upper endpoint as 9. Thus, the conditional PI is also [2, 9]. To compute the exact
PI, we note that the probability in (2) at L = 2 is H(20; 22, 400, 500) = .1485, and at L = 1 is
1 H(19; 21, 400, 500) = .0540. Thus, the lower endpoint of the exact PI is 1. The upper endpoint can
be similarly obtained as 10. Thus, exact PI is [1, 10], which is the widest among all PIs.
4.2 An Example for Poisson Prediction Intervals
We shall use the example given in Hahn and Chandra (1981) for illustrating the construction of
PIs for a Poisson distribution. The number of unscheduled shutdowns per year in a large population
of systems follows a Poisson distribution with mean . Suppose there were 24 shutdowns over a period
of 5 years, and we like to nd 90% PIs for the number of unscheduled shutdowns over the next year.
Here n = 5, x = 24 and m = 1. The maximum likelihood estimate of is given by

=
24
5
= 4.8.
Nelsons PI in (7) is computed as [4.8 3.95|, 4.8 + 3.95|] = [1, 8]. The new PI based on the joint
sampling approach given in (19) is [2, 9]. To compute the conditional PI, we note that m/(m + n) =
.1667 and
B(8; 24 + 8, .1667) = 0.9271 and B(9; 24 + 9, .1667) = 0.9621.
So 9 is the upper endpoint of the 90% conditional PI. Furthermore, B(a1; 24+a, .1667) = 0.0541 at a =
2 and B(a 1; 24 +a, .1667) = 0.0105 at a = 1, and so the lower endpoint of the 90% PI is 1. Thus,
the conditional PI is [1, 9]. To nd the lower endpoint of the exact PI using (14), we note that
n/(n + m) = 5/6 = .8333, and 1 B(23; 24 + 1, .8333) = .0628. It can be veried that L = 1 is the
smallest integer for which 1 B(23; 24 + L, .8333 > .05. Similarly, using (15), we can nd the upper
endpoint of the 90% exact PI as 9. Thus, the exact PI and the conditional PI are the same [1, 9]
whereas the new PI is [2, 9].
5. Concluding Remarks
It is now well-known that the classical exact methods for discrete distributions are often too con-
servative producing condence intervals that are unnecessarily wide or tests that are less powerful.
Examples include Clopper-Pearson (1934) condence interval for a binomial proportion, Fishers exact
test for comparing two proportions, Przyborowski and Wilenskis (1940) conditional test for the ratio
of two Poisson means, and Thomas and Garts (1977) method for estimating the odds ratio involving
two binomial distributions; see Agresti and Coull (1998), Agresti (1999) and Brown, Cai and Das
Gupta (2001). Many authors have recommended alternative approximate approaches for constructing
condence intervals with satisfactory coverage probabilities and good precision. In this article, we have
established similar results for constructing PIs for a binomial or Poisson distribution. Furthermore, we
showed that the conditional PIs are narrower than the exact PIs, and so the former PIs have shorter
17
expected width. The joint sampling approach produces PIs for both binomial and Poisson distributions
with good coverage properties and precisions. In terms of simplicity and accuracy, the new PIs are
preferable to others, and can be recommended for applications.
Acknowledgements The authors are grateful a reviewer for providing useful comments and sugges-
tions.
References
Aitchison, J., Dunsmore, I. R. 1980. Statistical Prediction Analysis. Cambridge University Press, Cambridge,
UK.
Agresti, A. 1999. On logit condence intervals for the odds ratio with small samples. Biometrics 55, 597-602.
Agresti, A., Coull, B. A. 1988. Approximate is better than exact for interval estimation of binomial propor-
tion. The American Statistician 52, 119-125.
Bain, L.J., Patel, J. K. 1993. Prediction intervals based on partial observations for some discrete distributions.
IEEE Transactions on Reliability 42, 459-463.
Brown, P. J. 1982. Multivariate calibration. Journal of the Royal Statistical Society, Ser. B 44, 287-321.
Brown L. D., Cai T., Das Gupta A. 2001. Interval estimation for a binomial proportion (with discussion).
Statistical Science 16, 101-133.
Clopper, C. J., Pearson E. S. 1934. The use of condence or ducial limits illustrated in the case of the
binomial. Biometrika 26, 404-413.
Davis, C. B., McNichols, R. J. 1987. One-sided intervals for at least p of m observations from a normal
population on each of r future occasions. Technometrics 29, 359-370.
Dunsmore, I. R. 1976. A note on faulkenberrys method of obtaining prediction intervals. Journal of the
American Statistical Association 71, 193-194.
Faulkenberry, G. D. 1973. A method of obtaining prediction intervals. Journal of the American Statistical
Association 68, 433-435.
Gibbons, R. D. 1987. Statistical prediction intervals for the evaluation of ground-water quality. Ground Water
25, 455-265.
Hahn, G. J., Meeker, W. Q. 1991. Statistical Intervals: A Guide to Practitioners, Wiley, New York.
Hahn, G. J., Chandra, R. 1981. Tolerance intervals for Poisson and binomial variables. Journal of Quality
Technology 13, 100-110.
Kn usel, L. 1994. The prediction problem as the dual form of the two-sample problem with applications to the
poisson and the binomial distribution. The American Statistician 48, 214-219.
Nelson, W. 1982. Applied Life Data Analysis. Wiley, New York.
Przyborowski, J., Wilenski, H. 1940. Homogeneity of results in testing samples from Poisson series. Biometrika
31, 313-323.
18
Thatcher, A. R. 1964. Relationships between bayesian and condence limits for prediction. Journal of the
Royal Statistical Society, Ser. B 26, 176-192.
Thomas, D. G. and Gart, J. J. 1977. A table of exact condence limits for dierences and ratios of two
proportions and their odds ratios. Journal of the American Statistical Association 72, 73-76.
Wang, H. 2008. Coverage probability of prediction intervals for discrete random variables. Computational
Statistics and Data Analysis 53, 17-26.
Wang, H. 2010. Closed form prediction intervals applied for disease counts. The American Statistician 64,
250-256.
Appendix A
We shall now show that, for the binomial case, the conditional PI is always included in or equal to the
exact PI. Let L
c
and L
e
be the lower endpoints of the 1 2 conditional PI and exact PI, respectively.
To show that L
c
is always greater than or equal to L
e
, we recall that L
c
is the largest integer so that
P(Y L
c
1[x + L
c
, m, n + m) . This is the probability of observing L
c
1 or fewer defects
from a sample of size x +L
c
from a lot of size n +m that contains m defects. This is the same as the
probability of observing at least x+1 non-defects in a sample of size x+L
c
from a lot of size n+m that
contains n non-defects. That is, P(Y L
c
1[x+L
c
, m, n+m) = P(X x+1[x+L
c
, n, n+m) .
Note that x must be the smallest integer for which the above inequality holds. Therefore,
P(X x + 1[x +L
c
, n, n +m) P(X x[x +L
c
, n, n +m) > .
Thus, L
c
satises the condition in (2), and so it is greater than or equal to L
e
. To show that L
c
> L
e
for some sample sizes, recall that L
e
is the smallest integer for which P(X x[x +L
e
, n, n +m) > .
Arguing as before, it is easy to see that P(X x[x+L
e
, n, n+m) = P(Y L
e
[x+L
e
, m, n+m) > .
Since L
e
is the smallest integer for which this inequality holds, P(Y L
e
1[x+L
e
1, m, n+m) .
As noted in Thatcher (1964), the hypergeometric cdf is a decreasing function of the sample size while
other quantities are xed, and so
P(Y L
e
1[x +L
e
, m, n +m) < .
However, L
e
is not necessarily the largest integer satisfying the above inequality, and so it could be
smaller than L
c
. In fact, there are many situations where L
c
> L
e
. For examples, when n = 40, m = 32
and x = 22, the 95% exact and conditional PIs intervals are [10, 25] and [11, 24], respectively; when
(n, m, x) = (36, 22, 18), they are [5, 17] and [6, 16], respectively.
We shall now show that the upper endpoint U
c
of a 1 2 conditional PI is included in the
corresponding exact PI. Recall that U
c
is the smallest integer for which
P(Y U
c
[x +U
c
, m, n +m) 1 . (20)
19
As argued earlier, the above probability is the same as the probability of observing at least x non-defects
in a sample of size x+U
c
when the number of non-defects is n. That is, P(Y U
c
[x+U
c
, m, n+m) =
P(X x[x + U
c
, n, n + m). Because U
c
is the smallest integer satisfying (20), x must be the largest
integer satisfying P(X x[x +U
c
, n, n +m) 1 . Thus, if U
c
satises (20), then
P(X x[x +U
c
, n, n +m) 1 P(X x 1[x +U
c
, n, n +m) , (21)
which implies that P(X x[x+U
c
, n, n+m) > . In other words, U
c
satises (3), and so it is included
in the exact PI. Furthermore, using the argument similar to the one given for the lower endpoint, it
can be shown that there are situations where U
c
could be strictly less than the upper endpoint of the
corresponding exact PI.
Appendix B
We here show that the JS-PI in (12) is included in or identical to Wangs PI determined by (8).
For a given (n, m, X), let Y
0
be a value that is in the JS-PI. To show that Y
0
is included in Wangs
PI, it is enough to show that if Y
0
satises (12) then it also satises [W(n, m, X, Y
0
)[ < z
1
, where
W(n, m, X, Y ) is dened in (8). Equivalently, we need to show that
X +Y
0
m+n
_
1
X +Y
0
m+n
_

X +Y
0
+c/2
m+n +c
_
1
X +Y
0
+c/2
m+n +c
_
,
where c = z
2
1
. Let d = X +Y
0
and N = m+n. It is easy to check that the above inequality holds if
and only if
d
2
dN +
N
2
4
0.
The above quadratic function of d attains its minimum at d = N/2, and the minimum value is zero.
Thus, the JS-PI is included in or identical to Wangs PI.